Currie, Iain Heriot-Watt University, Department of Actuarial Mathematics & Statistics Edinburgh EH14 4AS, UK
|
|
- Arlene Pearson
- 5 years ago
- Views:
Transcription
1 An Introduction to Generalized Linear Array Models Currie, Iain Heriot-Watt University, Department of Actuarial Mathematics & Statistics Edinburgh EH14 4AS, UK 1 Motivating example We use Swedish data downloaded from the Human Mortality Database (HMD) to motivate the need for a high speed, low storage method of smoothing. We suppose that the matrix Y = (y i,j ), i = 1,..., n a, j = 1,..., n y, contains the number of deaths in the j th year at the i th age. In our example, we consider ages x a = (x a,1,..., x a,na ) = (10,..., 90) and years x y = (x y,1,..., x y,ny ) = (,..., ). The matrix E = (e i,j ), n a n y, contains the corresponding exposed to risk, ie, the total time lived at the age x a,i in year x y,j. Let M = (m i,j ) = (log(y i,j /e i,j )), n a n y, be the matrix of the observed forces of mortality (measured on the log scale). The rows and columns of Y, E and M are indexed by age x a and year x y respectively. The left panel of Figure 1, a plot of the observed mortality surface, suggests the following: the force of mortality (1) has fallen steadily throughout the th century, (2) it increases steadily with age, except that (3) it increases then decreases rapidly around the age of twenty (4) it exhibited a very large systematic increase around the end of the first world war and (5) it is subject to random departure from these underlying patterns. It is natural to suppose that underlying the observed mortality surface in Figure 1 is a smooth mortality surface. We estimate this smooth surface with the P -spline system of Eilers & Marx (1996). Let {B a,1,..., B a,ca } be a B-spline basis of dimension c a defined along age and let B a = B a (x a ) = (B a,j (x a,i )), n a c a, be the resulting regression matrix. Similarly, let {B y,1,..., B y,cy } be a B-spline basis of dimension c y defined along year and let B y = B y (x y ) = (B y,j (x y,i )), n y c y, be the resulting regression matrix. Then a suitable model matrix for 2-d smoothing is given by the Kronecker product (1) B = B y B a, n a n y c a c y. The right panel of Figure 1 is a simplified plot of the underlying 2-d basis. Each of the c a c y regression coefficients is associated with the summit of a 2-d B-spline in the right panel of Figure 1. Thus, it is natural to think of these coefficients as arranged in a c a c y matrix, say Θ. The P -spline system now consists of choosing a rich basis in both age and year. Smoothness is ensured by penalizing adjacent coefficients in the rows and columns of Θ; the appropriate penalty matrix (Currie et al., 04) is (2) P = λ a I cy D ad a + λ y D yd y I ca, where I n is the identity matrix of size n, D a, (c a d a ) c a, and D y, (c y d y ) c y, are difference matrices of order d a and d y respectively (often we take d a = d y = 2). The difference matrices D a and D y penalize the coefficients in the columns and rows of Θ and λ a and λ y are the smoothing parameters in the age and year directions; notice that the P -spline system allows non-isotropic smoothing. Finally we suppose that the numbers of deaths y i,j are realizations of independent Poisson distributions with means µ i,j = e i,j φ i,j where φ i,j is the force of mortality at age x a,i in year x y,j. The standard approach to fitting this model is to vectorize the data and interpret the model as a penalized generalized linear model (PGLM). We have the following model structure and estimation algorithm Data: vectors y = vec Y, deaths, and e = vec E, exposures
2 Raw mortality surface 2d B spline basis B spline Figure 1: Left: Observed mortality surface for Swedish males. Right: Simplified 2-d B-spline basis. Model: a model matrix B = B y B a of B-splines, a parameter vector θ, θ = vec (Θ) and a link function (3) µ = E(y), log µ = log e + Bθ. Error distribution: Poisson. Algorithm: Penalized scoring algorithm (4) (B Wδ B + P ) ˆθ = B Wδ z where z = B θ + W 1 δ (y µ) is the working vector, is the penalty matrix, (2). Wδ is a diagonal matrix of weights and P This is a medium size problem: the length of y is = 8181 and a typical model matrix with c a = 15 and c y = is The smoothing parameters, λ a and λ y, must now be chosen in this PGLM framework. This is quite possible but we can see that, with a larger problem in higher dimensions where three or more smoothing parameters must be chosen, we may start to run into serious difficulties both in computational time and even in storage. From an aesthetic point of view there is something unnatural about the above structure. Our data and coefficients are matrices, and our model matrix is a product, yet the above analysis makes no use of these structures. We show in the next section that by adopting an array approach we can solve the storage problem and reduce computational time by up to orders of magnitude. 2 Generalized linear array models or GLAM The key formula follows from a well known property of Kronecker products (5) [B y B a ]θ, n a n y 1 B a ΘB y, n a n y, where indicates both sides have the same elements, although their dimensions are different. On the left we use the full model matrix, B = B y B a to evaluate Bθ; on the right, we operate on the matrix of coefficients first by B a and then by B y. This solves the storage problem. The number
3 of multiplications required to evaluate the right hand side is also very substantially smaller than is required on the left hand side. In a large problem this computational saving can be a factor of several orders of magnitude. We now give the model structure and estimation algorithm for the generalized linear array model or GLAM approach. In 2-dimensions an array is simply a matrix. Data: matrices Y, deaths, and E, exposures. Model: a model matrix B = B y B a of B-splines, a parameter matrix Θ and a link function (6) log E(Y ) = log E + B a ΘB y. Error distribution: Poisson. Algorithm: Penalized scoring algorithm (4) with (7) (8) Bθ, n a n y 1 B a ΘB y, n a n y, B W δ B, c a c y c a c y G(B a ) W G(B y ), c 2 a c 2 y, where W, n a n y, is the matrix of weights, ie, vec W = diag W δ. In (8), the function G( ) is the row-tensor function defined for any matrix X, n c, as (9) G(X) = [ X 1 c] [ 1 c X ], n c 2, where 1 c is a vector of 1 s of length c. We make two comments on the matrix form of this algorithm. First, (8) achieves the same sequential computation for the weighted inner product B W δ B in the scoring algorithm (4) as (6) achieves for the linear predictor. Second, the matrix forms on the righthand sides of (7) and (8) need to be rearranged into the corresponding vector forms on their respective left-hand sides. Details of how this is achieved are given in Currie et al, (06); here, we simply remark that such rearrangements are very efficient. In summary, GLAM is conceptually attractive (it takes advantage of both the data and the model structure), has a low footprint (a result of the sequential nature of the algorithm), is very fast, and generalizes to d-dimensions. The sequential nature of the GLAM algorithm was first described in Eilers et al. (06). The array nature of these algorithms was described in Currie et al. (06); the acronym GLAM was coined in this second paper. 3 Examples of GLAMs In the remainder of this paper we describe a number of applications of GLAM. We will describe the data, give the linear predictor in GLAM form and report some results.
4 Observed and fitted log mortality for age Observed and fitted log mortality for year 1950 Figure 2: Smoothed log mortality for Swedish males 3.1 Modelling mortality We continue with our introductory example on smoothing Swedish mortality data. We ignore for the moment the one-off behaviour at the end of the first world war and fit a simple smooth surface B a ΘB y. Smoothing parameters are chosen by minimizing the Bayesian Information Criterion (BIC). Figure 2 gives two cross-sections of the fitted surface. The fitted surface has successfully picked up the feature around age (often described as the accident hump ) but has failed to catch the feature in We will return to this point in section 3.3. There is an R-package, MortalitySmooth, (Camarda, 09) for fitting a 2-d mortality surface; overdispersion of the Poisson counts, a common feature of mortality data, is allowed for. The coding uses the GLAM algorithms. 3.2 Joint modelling of mortality by lives and amounts Mortality data in life insurance has some features of its own. Deaths are the number of claims on policies and exposure is the total time at risk; this is usually referred to as data by lives. Alternatively, deaths are the total amount claimed and exposure is the total amount at risk; this is usually referred to as data by amounts. Thus we have deaths and exposure matrices D l and E l on lives, and D a and E a on amounts. The force of mortality can then be computed by lives (as in the previous example) or by amounts. It is generally found that mortality by lives is heavier than mortality by amounts since those with better mortality have larger insured amounts; this observation is borne out in Figure 3. Actuaries are interested in forecasting mortality for the purpose of pricing and reserving of annuities and pensions. The penalty function allows forecasting (see Currie et al. (04) for details of how this is done). Here, with two measures of mortality we must take care that any such forecasts are consistent, ie, do not cross over in the future. We achieve this joint forecasting by taking B a ΘB y as the mortality surface by lives and B a ΘB y + B a Θ1 ny as the mortality surface by amounts. This is an additive GLAM with the property that the surface by amounts differs from the surface by lives by a constant amount in time; this constant is a smooth function of age. There are three smoothing parameters to be chosen (two for the component B a ΘB y and one for the component B a Θ1 ny. Figure 3 shows the results of fitting this model. See Currie et al. (04) and Djeundje and Currie (11) for further details.
5 Log(mortality) Amounts = 70 Lives Log(mortality) Lives Amounts = Figure 3: Smoothed and forecast log mortality by lives and amounts 3.3 Modelling mortality with period shocks We return to the Swedish mortality data and the problem with the one-off feature in Rather than use a model with the observed feature in 1918 built-in we propose a more general model in which each year can potentially have such a feature. We refer to such features as period shocks or simply as shocks. We assume that any period shocks are smooth functions of age; we leave the modelling process to identify them. Kirkby and Currie (10) proposed an additive GLAM (10) B a ΘB y + B a Θ which splits the mortality surface into a smooth 2-d surface and (possibly) a number of smooth shocks. This is a large computationally demanding model since each of n y years has c a coefficients. Thus, the three smoothing parameters must be chosen in the context of a very large model matrix; in Kirkby and Currie (10) the model matrix was Figure 4 shows the two components of the additive model (10). The smooth surface in the left panel of Figure 4 is much smoother than that fitted in section 3.1 since the one-off behaviour in 1918 is now modelled by a shock. The shock to the mortality surface in 1918 was caused by the Spanish Influenza pandemic which predominately affected the young. The 1918 shock and other smaller shocks can be seen in the right panel of Figure d density estimation GLAM can also be applied in some situations when the data do not obviously lie on a grid, and 2-d density estimation is a good example. We consider the classic 2-d data set on the waiting times between and durations of eruptions of the Old Faithful geyser. The left panel of Figure 5 shows the data from which it is evident that data falls into two main regions. We have 272 data points scattered over a rectangular region. We form a fine 2-d grid of counts. In the example we used waiting time by
6 Smooth Shocks Figure 4: Components of the shock model. Left: the smooth 2-d surface; right: the annual shocks duration bins of size 1 minute by 1 second which gives 217 = 130 bins. Our data then consists of the frequencies of observation in cells: we have 238 counts of 1, 17 of 2, and (98%) of 0. We apply 2-d P -spline smoothing with Poisson errors, log link, model matrix B w (x w ) B d (x d ) where x w and x d are the mid-points of the waiting time and duration bins respectively. Figure 5 shows the fitted contour surface and fitted density. Further details on density estimation with GLAM can be found in Eilers and Marx (06). Durban et al. (06) combine GLAM and mixed models to estimate multi-dimensional densities while Lambert and Eilers (06) show how GLAM and Bayesian methods can also be used here. 3.5 Other applications One important area where GLAM can be used with advantage is in spatio-temporal smoothing. Lee and Durban (11) give an example of smoothing ozone measurements. Data are located at scattered locations but are measured at monthly intervals. This gives this 3-dimensional problem (two space dimensions and one time dimension) a 2-dimensional GLAM structure where one GLAM dimension is space and the other GLAM dimension is time. For further details see Lee and Durban (11). We have emphasized the computational advantage of GLAM in this introductory paper. There is another important aspect: if data have an array structure and models of interest have a row and column structure then GLAM is the correct way to think about modelling. GLAM is more than a computational device, it is a structure for modelling. REFERENCES Camarda, C. G. (09). MortalitySmooth: Smoothing Poisson counts with P -splines. R-package: 1.0. Currie, I. D., Durban, M. and Eilers, P. H. C. (04). Smoothing and forecasting mortality rates. Statistical Modelling, 4, Currie, I. D., Durban, M. and Eilers, P. H. C. (06). Generalized linear array models with applications to multidimensional smoothing. Journal of the Royal Statistical Society, Series B, 68, Djeundje, V. A. B. and Currie, I. D. (11). Smoothing dispersed counts with applications to mortality data. Annals of Actuarial Science, 5, Durban, M., Currie, I. D. and Eilers, P. H. C. (06). Mixed models, array methods and multidimensional density estimation. Proceedings of 21st International Workshop on Statistical Modelling, Galway, Eilers, P. H. C., Currie, I. D. and Durban, M. (06). Fast and compact smoothing on large multidimen-
7 Normalized Density Normalized Density Duration (minutes): bin width = 1 sec Density Waiting time Duration Waiting time (minutes): bin width = 1 min Figure 5: Data and estimate of 2-d density of waiting and durations times of eruptions of the Old Faithful geyser sional grids. Computational Statistics and Data Analysis, 50, Eilers, P. H. C. and Marx, B. D. (1996). Flexible smoothing with B-splines and penalties. Statistical Science, 11, Eilers, P. H. C. and Marx, B. D. (06). Multidimensional density smoothing with P -splines. Proceedings of 21st International Workshop on Statistical Modelling, Galway, Kirkby, J. G. and Currie, I. D. (10). Smooth models of mortality with period shocks. Statistical Modelling, 10, Lambert, P. and Eilers, P. H. C. (06). Bayesian multi-dimensional density estimation with P -splines. Proceedings of 21st International Workshop on Statistical Modelling, Galway, Lee, D-J. and Durban, M. (11). Pspline ANOVA type interaction models for spatio temporal smoothing. Statistical Modelling, 11, ABSTRACT Data with an array structure are common in statistics (mortality tables and spatio-temporal data are two important examples). Such data often require smoothing to remove noise and estimate trends. One natural and attractive approach is to use penalized regression where (a) the basis for the regression is a Kronecker product of B-splines and (b) the penalty is a roughness penalty on regression coefficients; this is the P-spline approach of Eilers & Marx. However, such an approach is particularly susceptible to runaway problems with (a) storage and (b) computational time. Generalized linear array models (GLAM) were developed precisely to address both these issues. In a conventional GLM you store the model matrix and then fit the model. Unfortunately, with large amounts of data this model matrix can get rather large: computation and even storage can be a problem. In GLAM the model matrix is not stored; the GLAM algorithm works sequentially with the factors of the Kronecker product. Further, the GLAM algorithm is very fast and can be orders of magnitude quicker than the usual GLM approach in a large problem. In this paper we first describe the GLAM algorithms and then give an introduction to a range of applications. These applications include various models for smoothing and forecasting of mortality tables, density estimation and spatio-temporal smoothing.
GLAM An Introduction to Array Methods in Statistics
GLAM An Introduction to Array Methods in Statistics Iain Currie Heriot Watt University GLAM A Generalized Linear Array Model is a low-storage, high-speed, method for multidimensional smoothing, when data
More informationMultidimensional Density Smoothing with P-splines
Multidimensional Density Smoothing with P-splines Paul H.C. Eilers, Brian D. Marx 1 Department of Medical Statistics, Leiden University Medical Center, 300 RC, Leiden, The Netherlands (p.eilers@lumc.nl)
More informationP -spline ANOVA-type interaction models for spatio-temporal smoothing
P -spline ANOVA-type interaction models for spatio-temporal smoothing Dae-Jin Lee 1 and María Durbán 1 1 Department of Statistics, Universidad Carlos III de Madrid, SPAIN. e-mail: dae-jin.lee@uc3m.es and
More informationFlexible Spatio-temporal smoothing with array methods
Int. Statistical Inst.: Proc. 58th World Statistical Congress, 2011, Dublin (Session IPS046) p.849 Flexible Spatio-temporal smoothing with array methods Dae-Jin Lee CSIRO, Mathematics, Informatics and
More informationUsing P-splines to smooth two-dimensional Poisson data
1 Using P-splines to smooth two-dimensional Poisson data Maria Durbán 1, Iain Currie 2, Paul Eilers 3 17th IWSM, July 2002. 1 Dept. Statistics and Econometrics, Universidad Carlos III de Madrid, Spain.
More informationSpace-time modelling of air pollution with array methods
Space-time modelling of air pollution with array methods Dae-Jin Lee Royal Statistical Society Conference Edinburgh 2009 D.-J. Lee (Uc3m) GLAM: Array methods in Statistics RSS 09 - Edinburgh # 1 Motivation
More informationA Hierarchical Perspective on Lee-Carter Models
A Hierarchical Perspective on Lee-Carter Models Paul Eilers Leiden University Medical Centre L-C Workshop, Edinburgh 24 The vantage point Previous presentation: Iain Currie led you upward From Glen Gumbel
More informationRecovering Indirect Information in Demographic Applications
Recovering Indirect Information in Demographic Applications Jutta Gampe Abstract In many demographic applications the information of interest can only be estimated indirectly. Modelling events and rates
More informationArray methods in statistics with applications to the modelling and forecasting of mortality. James Gavin Kirkby
Array methods in statistics with applications to the modelling and forecasting of mortality James Gavin Kirkby Submitted for the degree of Doctor of Philosophy Heriot-Watt University School of Mathematical
More informationSmoothing Age-Period-Cohort models with P -splines: a mixed model approach
Smoothing Age-Period-Cohort models with P -splines: a mixed model approach Running headline: Smooth Age-Period-Cohort models I D Currie, Department of Actuarial Mathematics and Statistics, and the Maxwell
More informationSmoothing Age-Period-Cohort models with P -splines: a mixed model approach
Smoothing Age-Period-Cohort models with P -splines: a mixed model approach Running headline: Smooth Age-Period-Cohort models I D Currie, Department of Actuarial Mathematics and Statistics, and the Maxwell
More informationINSTITUTE AND FACULTY OF ACTUARIES. Curriculum 2019 SPECIMEN SOLUTIONS
INSTITUTE AND FACULTY OF ACTUARIES Curriculum 09 SPECIMEN SOLUTIONS Subject CSA Risk Modelling and Survival Analysis Institute and Faculty of Actuaries Sample path A continuous time, discrete state process
More informationMULTIDIMENSIONAL COVARIATE EFFECTS IN SPATIAL AND JOINT EXTREMES
MULTIDIMENSIONAL COVARIATE EFFECTS IN SPATIAL AND JOINT EXTREMES Philip Jonathan, Kevin Ewans, David Randell, Yanyun Wu philip.jonathan@shell.com www.lancs.ac.uk/ jonathan Wave Hindcasting & Forecasting
More informationBayesian Nonparametric Regression for Diabetes Deaths
Bayesian Nonparametric Regression for Diabetes Deaths Brian M. Hartman PhD Student, 2010 Texas A&M University College Station, TX, USA David B. Dahl Assistant Professor Texas A&M University College Station,
More informationBayesian covariate models in extreme value analysis
Bayesian covariate models in extreme value analysis David Randell, Philip Jonathan, Kathryn Turnbull, Mathew Jones EVA 2015 Ann Arbor Copyright 2015 Shell Global Solutions (UK) EVA 2015 Ann Arbor June
More informationTowards a Regression using Tensors
February 27, 2014 Outline Background 1 Background Linear Regression Tensorial Data Analysis 2 Definition Tensor Operation Tensor Decomposition 3 Model Attention Deficit Hyperactivity Disorder Data Analysis
More informationIntroduction. Chapter 1
Chapter 1 Introduction In this book we will be concerned with supervised learning, which is the problem of learning input-output mappings from empirical data (the training dataset). Depending on the characteristics
More informationSpatial bias modeling with application to assessing remotely-sensed aerosol as a proxy for particulate matter
Spatial bias modeling with application to assessing remotely-sensed aerosol as a proxy for particulate matter Chris Paciorek Department of Biostatistics Harvard School of Public Health application joint
More informationMortality projection based on the Wang transform
Mortality projection based on the Wang transform Piet de Jong and Claymore Marshall Department of Actuarial Studies, Macquarie University, NSW 219, Australia. Email: piet.dejong@mq.edu.au Abstract A new
More informationOn the Importance of Dispersion Modeling for Claims Reserving: Application of the Double GLM Theory
On the Importance of Dispersion Modeling for Claims Reserving: Application of the Double GLM Theory Danaïl Davidov under the supervision of Jean-Philippe Boucher Département de mathématiques Université
More informationStatistics 203: Introduction to Regression and Analysis of Variance Course review
Statistics 203: Introduction to Regression and Analysis of Variance Course review Jonathan Taylor - p. 1/?? Today Review / overview of what we learned. - p. 2/?? General themes in regression models Specifying
More informationSTA216: Generalized Linear Models. Lecture 1. Review and Introduction
STA216: Generalized Linear Models Lecture 1. Review and Introduction Let y 1,..., y n denote n independent observations on a response Treat y i as a realization of a random variable Y i In the general
More informationEXPERIENCE-BASED LONGEVITY ASSESSMENT
EXPERIENCE-BASED LONGEVITY ASSESSMENT ERMANNO PITACCO Università di Trieste ermanno.pitacco@econ.units.it p. 1/55 Agenda Introduction Stochastic modeling: the process risk Uncertainty in future mortality
More informationLecture 8. Poisson models for counts
Lecture 8. Poisson models for counts Jesper Rydén Department of Mathematics, Uppsala University jesper.ryden@math.uu.se Statistical Risk Analysis Spring 2014 Absolute risks The failure intensity λ(t) describes
More informationParameter Estimation in the Spatio-Temporal Mixed Effects Model Analysis of Massive Spatio-Temporal Data Sets
Parameter Estimation in the Spatio-Temporal Mixed Effects Model Analysis of Massive Spatio-Temporal Data Sets Matthias Katzfuß Advisor: Dr. Noel Cressie Department of Statistics The Ohio State University
More informationLinear Regression Models P8111
Linear Regression Models P8111 Lecture 25 Jeff Goldsmith April 26, 2016 1 of 37 Today s Lecture Logistic regression / GLMs Model framework Interpretation Estimation 2 of 37 Linear regression Course started
More informationLog-Density Estimation with Application to Approximate Likelihood Inference
Log-Density Estimation with Application to Approximate Likelihood Inference Martin Hazelton 1 Institute of Fundamental Sciences Massey University 19 November 2015 1 Email: m.hazelton@massey.ac.nz WWPMS,
More informationAppendix A. Numeric example of Dimick Staiger Estimator and comparison between Dimick-Staiger Estimator and Hierarchical Poisson Estimator
Appendix A. Numeric example of Dimick Staiger Estimator and comparison between Dimick-Staiger Estimator and Hierarchical Poisson Estimator As described in the manuscript, the Dimick-Staiger (DS) estimator
More informationCurve Fitting Re-visited, Bishop1.2.5
Curve Fitting Re-visited, Bishop1.2.5 Maximum Likelihood Bishop 1.2.5 Model Likelihood differentiation p(t x, w, β) = Maximum Likelihood N N ( t n y(x n, w), β 1). (1.61) n=1 As we did in the case of the
More informationGeneralized Linear Models
Generalized Linear Models Advanced Methods for Data Analysis (36-402/36-608 Spring 2014 1 Generalized linear models 1.1 Introduction: two regressions So far we ve seen two canonical settings for regression.
More informationGeneralized Linear Models. Kurt Hornik
Generalized Linear Models Kurt Hornik Motivation Assuming normality, the linear model y = Xβ + e has y = β + ε, ε N(0, σ 2 ) such that y N(μ, σ 2 ), E(y ) = μ = β. Various generalizations, including general
More informationPricing and Risk Analysis of a Long-Term Care Insurance Contract in a non-markov Multi-State Model
Pricing and Risk Analysis of a Long-Term Care Insurance Contract in a non-markov Multi-State Model Quentin Guibert Univ Lyon, Université Claude Bernard Lyon 1, ISFA, Laboratoire SAF EA2429, F-69366, Lyon,
More informationNew Statistical Methods That Improve on MLE and GLM Including for Reserve Modeling GARY G VENTER
New Statistical Methods That Improve on MLE and GLM Including for Reserve Modeling GARY G VENTER MLE Going the Way of the Buggy Whip Used to be gold standard of statistical estimation Minimum variance
More informationSparse Nonparametric Density Estimation in High Dimensions Using the Rodeo
Outline in High Dimensions Using the Rodeo Han Liu 1,2 John Lafferty 2,3 Larry Wasserman 1,2 1 Statistics Department, 2 Machine Learning Department, 3 Computer Science Department, Carnegie Mellon University
More informationSpacetime models in R-INLA. Elias T. Krainski
Spacetime models in R-INLA Elias T. Krainski 2 Outline Separable space-time models Infant mortality in Paraná PM-10 concentration in Piemonte, Italy 3 Multivariate dynamic regression model y t : n observations
More informationBayesian Hierarchical Models
Bayesian Hierarchical Models Gavin Shaddick, Millie Green, Matthew Thomas University of Bath 6 th - 9 th December 2016 1/ 34 APPLICATIONS OF BAYESIAN HIERARCHICAL MODELS 2/ 34 OUTLINE Spatial epidemiology
More informationMissing Data Issues in the Studies of Neurodegenerative Disorders: the Methodology
Missing Data Issues in the Studies of Neurodegenerative Disorders: the Methodology Sheng Luo, PhD Associate Professor Department of Biostatistics & Bioinformatics Duke University Medical Center sheng.luo@duke.edu
More informationVariable Selection and Model Choice in Survival Models with Time-Varying Effects
Variable Selection and Model Choice in Survival Models with Time-Varying Effects Boosting Survival Models Benjamin Hofner 1 Department of Medical Informatics, Biometry and Epidemiology (IMBE) Friedrich-Alexander-Universität
More informationConditional Transformation Models
IFSPM Institut für Sozial- und Präventivmedizin Conditional Transformation Models Or: More Than Means Can Say Torsten Hothorn, Universität Zürich Thomas Kneib, Universität Göttingen Peter Bühlmann, Eidgenössische
More informationSelection of Smoothing Parameter for One-Step Sparse Estimates with L q Penalty
Journal of Data Science 9(2011), 549-564 Selection of Smoothing Parameter for One-Step Sparse Estimates with L q Penalty Masaru Kanba and Kanta Naito Shimane University Abstract: This paper discusses the
More informationLecture 2: Univariate Time Series
Lecture 2: Univariate Time Series Analysis: Conditional and Unconditional Densities, Stationarity, ARMA Processes Prof. Massimo Guidolin 20192 Financial Econometrics Spring/Winter 2017 Overview Motivation:
More informationLattice Data. Tonglin Zhang. Spatial Statistics for Point and Lattice Data (Part III)
Title: Spatial Statistics for Point Processes and Lattice Data (Part III) Lattice Data Tonglin Zhang Outline Description Research Problems Global Clustering and Local Clusters Permutation Test Spatial
More informationRegularization in Cox Frailty Models
Regularization in Cox Frailty Models Andreas Groll 1, Trevor Hastie 2, Gerhard Tutz 3 1 Ludwig-Maximilians-Universität Munich, Department of Mathematics, Theresienstraße 39, 80333 Munich, Germany 2 University
More informationGLM I An Introduction to Generalized Linear Models
GLM I An Introduction to Generalized Linear Models CAS Ratemaking and Product Management Seminar March Presented by: Tanya D. Havlicek, ACAS, MAAA ANTITRUST Notice The Casualty Actuarial Society is committed
More informationDynamic sequential analysis of careers
Dynamic sequential analysis of careers Fulvia Pennoni Department of Statistics and Quantitative Methods University of Milano-Bicocca http://www.statistica.unimib.it/utenti/pennoni/ Email: fulvia.pennoni@unimib.it
More informationModel checking overview. Checking & Selecting GAMs. Residual checking. Distribution checking
Model checking overview Checking & Selecting GAMs Simon Wood Mathematical Sciences, University of Bath, U.K. Since a GAM is just a penalized GLM, residual plots should be checked exactly as for a GLM.
More informationEstimating the term structure of mortality
Insurance: Mathematics and Economics 42 (2008) 492 504 www.elsevier.com/locate/ime Estimating the term structure of mortality Norbert Hári a, Anja De Waegenaere b,, Bertrand Melenberg b,a, Theo E. Nijman
More informationLISA Short Course Series Generalized Linear Models (GLMs) & Categorical Data Analysis (CDA) in R. Liang (Sally) Shan Nov. 4, 2014
LISA Short Course Series Generalized Linear Models (GLMs) & Categorical Data Analysis (CDA) in R Liang (Sally) Shan Nov. 4, 2014 L Laboratory for Interdisciplinary Statistical Analysis LISA helps VT researchers
More informationComputational methods for mixed models
Computational methods for mixed models Douglas Bates Department of Statistics University of Wisconsin Madison March 27, 2018 Abstract The lme4 package provides R functions to fit and analyze several different
More informationGeneralized Linear Models. Last time: Background & motivation for moving beyond linear
Generalized Linear Models Last time: Background & motivation for moving beyond linear regression - non-normal/non-linear cases, binary, categorical data Today s class: 1. Examples of count and ordered
More informationPredictive spatio-temporal models for spatially sparse environmental data. Umeå University
Seminar p.1/28 Predictive spatio-temporal models for spatially sparse environmental data Xavier de Luna and Marc G. Genton xavier.deluna@stat.umu.se and genton@stat.ncsu.edu http://www.stat.umu.se/egna/xdl/index.html
More informationx y = 1, 2x y + z = 2, and 3w + x + y + 2z = 0
Section. Systems of Linear Equations The equations x + 3 y =, x y + z =, and 3w + x + y + z = 0 have a common feature: each describes a geometric shape that is linear. Upon rewriting the first equation
More informationEruptions of the Old Faithful geyser
Eruptions of the Old Faithful geyser geyser2.ps Topics covered: Bimodal distribution. Data summary. Examination of univariate data. Identifying subgroups. Prediction. Key words: Boxplot. Histogram. Mean.
More informationStatistical foundations
Statistical foundations Michael K. Tippett International Research Institute for Climate and Societ The Earth Institute, Columbia Universit ERFS Climate Predictabilit Tool Training Workshop Ma 4-9, 29 Ideas
More informationExtending clustered point process-based rainfall models to a non-stationary climate
Extending clustered point process-based rainfall models to a non-stationary climate Jo Kaczmarska 1, 2 Valerie Isham 2 Paul Northrop 2 1 Risk Management Solutions 2 Department of Statistical Science, University
More informationSurvival Analysis I (CHL5209H)
Survival Analysis Dalla Lana School of Public Health University of Toronto olli.saarela@utoronto.ca January 7, 2015 31-1 Literature Clayton D & Hills M (1993): Statistical Models in Epidemiology. Not really
More informationOutline of GLMs. Definitions
Outline of GLMs Definitions This is a short outline of GLM details, adapted from the book Nonparametric Regression and Generalized Linear Models, by Green and Silverman. The responses Y i have density
More informationNow consider the case where E(Y) = µ = Xβ and V (Y) = σ 2 G, where G is diagonal, but unknown.
Weighting We have seen that if E(Y) = Xβ and V (Y) = σ 2 G, where G is known, the model can be rewritten as a linear model. This is known as generalized least squares or, if G is diagonal, with trace(g)
More informationOn prediction and density estimation Peter McCullagh University of Chicago December 2004
On prediction and density estimation Peter McCullagh University of Chicago December 2004 Summary Having observed the initial segment of a random sequence, subsequent values may be predicted by calculating
More informationTriangles in Life and Casualty
Triangles in Life and Casualty Gary G. Venter, Guy Carpenter LLC gary.g.venter@guycarp.com Abstract The loss development triangles in casualty insurance have a similar setup to the mortality datasets used
More informationSTAT 135 Lab 11 Tests for Categorical Data (Fisher s Exact test, χ 2 tests for Homogeneity and Independence) and Linear Regression
STAT 135 Lab 11 Tests for Categorical Data (Fisher s Exact test, χ 2 tests for Homogeneity and Independence) and Linear Regression Rebecca Barter April 20, 2015 Fisher s Exact Test Fisher s Exact Test
More informationSTATS 200: Introduction to Statistical Inference. Lecture 29: Course review
STATS 200: Introduction to Statistical Inference Lecture 29: Course review Course review We started in Lecture 1 with a fundamental assumption: Data is a realization of a random process. The goal throughout
More informationCohort Effect Structure in the Lee-Carter Residual Term. Naoki Sunamoto, FIAJ. Fukoku Mutual Life Insurance Company
Cohort Effect Structure in the Lee-Carter Residual Term Naoki Sunamoto, FIAJ Fukoku Mutual Life Insurance Company 2-2 Uchisaiwaicho 2-chome, Chiyoda-ku, Tokyo, 100-0011, Japan Tel: +81-3-3593-7445, Fax:
More informationHierarchical Nearest-Neighbor Gaussian Process Models for Large Geo-statistical Datasets
Hierarchical Nearest-Neighbor Gaussian Process Models for Large Geo-statistical Datasets Abhirup Datta 1 Sudipto Banerjee 1 Andrew O. Finley 2 Alan E. Gelfand 3 1 University of Minnesota, Minneapolis,
More information4 Testing Hypotheses. 4.1 Tests in the regression setting. 4.2 Non-parametric testing of survival between groups
4 Testing Hypotheses The next lectures will look at tests, some in an actuarial setting, and in the last subsection we will also consider tests applied to graduation 4 Tests in the regression setting )
More informationEcon 2148, fall 2017 Gaussian process priors, reproducing kernel Hilbert spaces, and Splines
Econ 2148, fall 2017 Gaussian process priors, reproducing kernel Hilbert spaces, and Splines Maximilian Kasy Department of Economics, Harvard University 1 / 37 Agenda 6 equivalent representations of the
More informationModels for Underreporting: A Bernoulli Sampling Approach for Reported Counts
AUSTRIAN JOURNAL OF STATISTICS Volume 40 (2011), Number 1 & 2, 85-92 Models for Underreporting: A Bernoulli Sampling Approach for Reported Counts Gerhard Neubauer 1, Gordana Djuraš 1 and Herwig Friedl
More informationKneib, Fahrmeir: Supplement to "Structured additive regression for categorical space-time data: A mixed model approach"
Kneib, Fahrmeir: Supplement to "Structured additive regression for categorical space-time data: A mixed model approach" Sonderforschungsbereich 386, Paper 43 (25) Online unter: http://epub.ub.uni-muenchen.de/
More informationCourse 4 Solutions November 2001 Exams
Course 4 Solutions November 001 Exams November, 001 Society of Actuaries Question #1 From the Yule-Walker equations: ρ φ + ρφ 1 1 1. 1 1+ ρ ρφ φ Substituting the given quantities yields: 0.53 φ + 0.53φ
More informationChapter 3. Introduction to Linear Correlation and Regression Part 3
Tuesday, December 12, 2000 Ch3 Intro Correlation Pt 3 Page: 1 Richard Lowry, 1999-2000 All rights reserved. Chapter 3. Introduction to Linear Correlation and Regression Part 3 Regression The appearance
More informationSparse Linear Systems. Iterative Methods for Sparse Linear Systems. Motivation for Studying Sparse Linear Systems. Partial Differential Equations
Sparse Linear Systems Iterative Methods for Sparse Linear Systems Matrix Computations and Applications, Lecture C11 Fredrik Bengzon, Robert Söderlund We consider the problem of solving the linear system
More informationSTA 216: GENERALIZED LINEAR MODELS. Lecture 1. Review and Introduction. Much of statistics is based on the assumption that random
STA 216: GENERALIZED LINEAR MODELS Lecture 1. Review and Introduction Much of statistics is based on the assumption that random variables are continuous & normally distributed. Normal linear regression
More informationGeneralized linear mixed models (GLMMs) for dependent compound risk models
Generalized linear mixed models (GLMMs) for dependent compound risk models Emiliano A. Valdez, PhD, FSA joint work with H. Jeong, J. Ahn and S. Park University of Connecticut Seminar Talk at Yonsei University
More informationGARY G. VENTER 1. CHAIN LADDER VARIANCE FORMULA
DISCUSSION OF THE MEAN SQUARE ERROR OF PREDICTION IN THE CHAIN LADDER RESERVING METHOD BY GARY G. VENTER 1. CHAIN LADDER VARIANCE FORMULA For a dozen or so years the chain ladder models of Murphy and Mack
More informationModelling trends in the ocean wave climate for dimensioning of ships
Modelling trends in the ocean wave climate for dimensioning of ships STK1100 lecture, University of Oslo Erik Vanem Motivation and background 2 Ocean waves and maritime safety Ships and other marine structures
More informationReserving for multiple excess layers
Reserving for multiple excess layers Ben Zehnwirth and Glen Barnett Abstract Patterns and changing trends among several excess-type layers on the same business tend to be closely related. The changes in
More informationStochastic Modelling Unit 1: Markov chain models
Stochastic Modelling Unit 1: Markov chain models Russell Gerrard and Douglas Wright Cass Business School, City University, London June 2004 Contents of Unit 1 1 Stochastic Processes 2 Markov Chains 3 Poisson
More informationAnalysis Methods for Supersaturated Design: Some Comparisons
Journal of Data Science 1(2003), 249-260 Analysis Methods for Supersaturated Design: Some Comparisons Runze Li 1 and Dennis K. J. Lin 2 The Pennsylvania State University Abstract: Supersaturated designs
More informationRate Maps and Smoothing
Rate Maps and Smoothing Luc Anselin Spatial Analysis Laboratory Dept. Agricultural and Consumer Economics University of Illinois, Urbana-Champaign http://sal.agecon.uiuc.edu Outline Mapping Rates Risk
More informationTechnical Vignette 5: Understanding intrinsic Gaussian Markov random field spatial models, including intrinsic conditional autoregressive models
Technical Vignette 5: Understanding intrinsic Gaussian Markov random field spatial models, including intrinsic conditional autoregressive models Christopher Paciorek, Department of Statistics, University
More informationDistortion Risk Measures: Coherence and Stochastic Dominance
Distortion Risk Measures: Coherence and Stochastic Dominance Dr. Julia L. Wirch Dr. Mary R. Hardy Dept. of Actuarial Maths Dept. of Statistics and and Statistics Actuarial Science Heriot-Watt University
More informationModelling Survival Data using Generalized Additive Models with Flexible Link
Modelling Survival Data using Generalized Additive Models with Flexible Link Ana L. Papoila 1 and Cristina S. Rocha 2 1 Faculdade de Ciências Médicas, Dep. de Bioestatística e Informática, Universidade
More informationVarieties of Count Data
CHAPTER 1 Varieties of Count Data SOME POINTS OF DISCUSSION What are counts? What are count data? What is a linear statistical model? What is the relationship between a probability distribution function
More informationResearch Projects. Hanxiang Peng. March 4, Department of Mathematical Sciences Indiana University-Purdue University at Indianapolis
Hanxiang Department of Mathematical Sciences Indiana University-Purdue University at Indianapolis March 4, 2009 Outline Project I: Free Knot Spline Cox Model Project I: Free Knot Spline Cox Model Consider
More informationTemporal Multi-View Inconsistency Detection for Network Traffic Analysis
WWW 15 Florence, Italy Temporal Multi-View Inconsistency Detection for Network Traffic Analysis Houping Xiao 1, Jing Gao 1, Deepak Turaga 2, Long Vu 2, and Alain Biem 2 1 Department of Computer Science
More informationLinear model A linear model assumes Y X N(µ(X),σ 2 I), And IE(Y X) = µ(x) = X β, 2/52
Statistics for Applications Chapter 10: Generalized Linear Models (GLMs) 1/52 Linear model A linear model assumes Y X N(µ(X),σ 2 I), And IE(Y X) = µ(x) = X β, 2/52 Components of a linear model The two
More informationBAYESIAN DECISION THEORY
Last updated: September 17, 2012 BAYESIAN DECISION THEORY Problems 2 The following problems from the textbook are relevant: 2.1 2.9, 2.11, 2.17 For this week, please at least solve Problem 2.3. We will
More informationPenalized Loss functions for Bayesian Model Choice
Penalized Loss functions for Bayesian Model Choice Martyn International Agency for Research on Cancer Lyon, France 13 November 2009 The pure approach For a Bayesian purist, all uncertainty is represented
More informationChain ladder with random effects
Greg Fry Consulting Actuaries Sydney Australia Astin Colloquium, The Hague 21-24 May 2013 Overview Fixed effects models Families of chain ladder models Maximum likelihood estimators Random effects models
More informationGeneralized Elastic Net Regression
Abstract Generalized Elastic Net Regression Geoffroy MOURET Jean-Jules BRAULT Vahid PARTOVINIA This work presents a variation of the elastic net penalization method. We propose applying a combined l 1
More informationUNIVERSITY OF MASSACHUSETTS Department of Mathematics and Statistics Basic Exam - Applied Statistics Thursday, August 30, 2018
UNIVERSITY OF MASSACHUSETTS Department of Mathematics and Statistics Basic Exam - Applied Statistics Thursday, August 30, 2018 Work all problems. 60 points are needed to pass at the Masters Level and 75
More informationMODULE 6 LOGISTIC REGRESSION. Module Objectives:
MODULE 6 LOGISTIC REGRESSION Module Objectives: 1. 147 6.1. LOGIT TRANSFORMATION MODULE 6. LOGISTIC REGRESSION Logistic regression models are used when a researcher is investigating the relationship between
More informationChapter 5: Generalized Linear Models
w w w. I C A 0 1 4. o r g Chapter 5: Generalized Linear Models b Curtis Gar Dean, FCAS, MAAA, CFA Ball State Universit: Center for Actuarial Science and Risk Management M Interest in Predictive Modeling
More informationLoss Estimation using Monte Carlo Simulation
Loss Estimation using Monte Carlo Simulation Tony Bellotti, Department of Mathematics, Imperial College London Credit Scoring and Credit Control Conference XV Edinburgh, 29 August to 1 September 2017 Motivation
More informationCS281A/Stat241A Lecture 17
CS281A/Stat241A Lecture 17 p. 1/4 CS281A/Stat241A Lecture 17 Factor Analysis and State Space Models Peter Bartlett CS281A/Stat241A Lecture 17 p. 2/4 Key ideas of this lecture Factor Analysis. Recall: Gaussian
More informationBivariate Weibull-power series class of distributions
Bivariate Weibull-power series class of distributions Saralees Nadarajah and Rasool Roozegar EM algorithm, Maximum likelihood estimation, Power series distri- Keywords: bution. Abstract We point out that
More informationBrent Coull, Petros Koutrakis, Joel Schwartz, Itai Kloog, Antonella Zanobetti, Joseph Antonelli, Ander Wilson, Jeremiah Zhu.
Harvard ACE Project 2: Air Pollutant Mixtures in Eastern Massachusetts: Spatial Multi-resolution Analysis of Trends, Effects of Modifiable Factors, Climate, and Particle-induced Mortality Brent Coull,
More informationM. Matrices and Linear Algebra
M. Matrices and Linear Algebra. Matrix algebra. In section D we calculated the determinants of square arrays of numbers. Such arrays are important in mathematics and its applications; they are called matrices.
More informationS.R.S Varadhan by Professor Tom Louis Lindstrøm
S.R.S Varadhan by Professor Tom Louis Lindstrøm Srinivasa S. R. Varadhan was born in Madras (Chennai), India in 1940. He got his B. Sc. from Presidency College in 1959 and his Ph.D. from the Indian Statistical
More informationChecking the Poisson assumption in the Poisson generalized linear model
Checking the Poisson assumption in the Poisson generalized linear model The Poisson regression model is a generalized linear model (glm) satisfying the following assumptions: The responses y i are independent
More information